Current Server Time: 23:02 (Central Europe)

#apertus IRC Channel Logs

2024/04/22

Timezone: UTC


06:07
BogdanXOR
left the channel
06:48
BogdanXOR
joined the channel
07:19
BogdanXOR
left the channel
07:49
BogdanXOR
joined the channel
07:54
BogdanXOR
left the channel
08:53
BogdanXOR
joined the channel
08:58
BogdanXOR
left the channel
09:25
gecim21
joined the channel
09:49
BogdanXOR
joined the channel
09:54
BogdanXOR
left the channel
10:05
BogdanXOR
joined the channel
11:12
gecim21
left the channel
17:00
Bertl
MEETING TIME!
17:00
Bertl
who is here and wants to report something?
17:00
Bertl
se6astian is still on the road and might join later.
17:03
Bertl
not much to report from my side this week, mostly planning and testing for the upcoming designs.
17:06
Bertl
okay then ...
17:06
Bertl
MEETING CONCLUDED!
17:07
Bertl
feel free to add your reports later ...
20:26
BogdanXOR
left the channel
20:54
BogdanXOR
joined the channel
21:00
BogdanXOR
left the channel
21:12
BogdanXOR
joined the channel
21:17
BogdanXOR
left the channel
21:29
BogdanXOR
joined the channel
21:34
BogdanXOR
left the channel
21:47
BogdanXOR
joined the channel
21:54
BogdanXOR
left the channel
22:18
BogdanXOR
joined the channel
22:24
anuejn
I do not have a lot to report
22:24
anuejn
just one random experiment:
22:25
anuejn
I tried to implement wavelet transform for compression using pytorch
22:25
anuejn
to accelerate it using hardware accellerators for convolutional neuronal networks
22:25
anuejn
as wavelet transform is basically a convolution
22:26
anuejn
and the rockchip has dedicated hardware for that
22:27
anuejn
on my M1 npu I got over 400fps
22:27
anuejn
for a 3 level haar cineform wavelet
22:27
anuejn
(just the transform, no quantization or bit packing)
22:28
anuejn
but I still found it quite interesting that it is indeed possible to use that kind of hardware for that without much hassle
22:28
anuejn
additionally vup convinced me, that a mipi plugin module with the lp lines equal for all lanes would probably work with some limitations
22:29
anuejn
that means that we could collaborate to build such a plugin, Bertl
22:31
anuejn
based on either the xilinx or the lattice appnote for a resistor phy and some (possibly 74xx ic) for buffering the lp signals
22:33
anuejn
(the main limitation is that we cannot send arbitrary width frames)
22:33
anuejn
vup calculated, that e.g. 4096 is impossible but 4100 would work
22:34
anuejn
and we could easily use the extra width to transport metadata
22:34
anuejn
which would be nice in any case